Concierge Architecture & Specs
7 dedicated AI agents (4 event-driven + 3 operational-input). Zero-polling real-time event-bus reaction in single-digit ms. Strict payload validation. An audit log entry is created before every invocation. Honest no-op design: returns a plain-language reason when no AI provider is wired, rather than fabricating success.
Capabilities & Orchestration
Event-driven agents (4)
VIP Welcome / Lead Enrichment / Sentiment Guard / Loyalty Liaison
Operational-input agents (3)
Chief of Staff / Assistant / Travel
Audit log
every agent invocation creates an entry before it runs, which is later updated with status, result, and a plain-language failure reason if something goes wrong
Admin helpers
Agent admin actions (list, retry, dismiss, configure thresholds) and budget alerts (per-agent cost caps)
Event Reaction Time
Single-digit ms via the platform's real-time event bus
Polling Overhead
0% — purely event-driven for the 4 event-driven agents; the 3 operational-input agents are cron-designed but invoked via workflow action today
Data Validation
Strict schema validation on every event payload and every agent input
Honest No-Op
When AI providers aren't configured, the entry records a plain-language reason (no provider connected / provider timed out / response failed validation) rather than fabricating a success
Audit Trail
An audit entry is created before the AI is ever called, then updated with the completed result and an AI-generated flag on success
Cross-Module Idempotency
The same event re-delivered does NOT re-fire the agent if an audit entry already exists for that event/record combo
Shared Agent Framework
All 3 operational-input agents run through the same internal framework, so the AI call, response validation, and audit logging behave consistently across all of them
Confidence Threshold
Operational-input agents flag outputs below configured threshold (default 60) for manual review rather than auto-actioning
Graceful Degradation
If a tenant hasn't installed the Inbox module, the VIP Welcome Agent skips the email-drafting phase but still completes CRM tagging + Task generation
1. Zero-Polling Event Architecture
The fundamental flaw of external automation tools (like Zapier or Make) is their reliance on REST API polling or brittle HTTP webhooks. - Native event bus: Concierge agents are wired directly into the platform's real-time event bus. When an order is paid by the Commerce module, the payload hits the Concierge agent in single-digit milliseconds without ever traversing the public internet. - Strict payload validation: Event payloads are inherently weakly typed at runtime. Every Concierge agent validates its input against a strict schema at the entry point before executing any cross-module logic.
2. The Extensible Agent Pattern
Concierge does not own business data; it is purely a logic orchestrator, built on a consistent pattern that makes it straightforward to add new agents over time. - Consistent agent shape: Every agent — built-in or added later — follows the same pattern: watch for a trigger, validate the input, call the AI, orchestrate the resulting actions, and write to the shared audit log. - Graceful Degradation: Because Concierge orchestrates across the OS, it must handle diverse tenant deployments. If a tenant hasn't installed the Inbox module, the VIP Welcome Agent gracefully skips the email-drafting phase but successfully completes the CRM tagging and Task generation. - Loyalty Liaison Specifics: Despite its name, the Loyalty Liaison Agent reacts to hotel check-ins rather than commerce orders — it welcomes returning Platinum or Diamond loyalty-tier guests the moment they check in.
3. The Operational-Input Agent Pattern
| Agent | Trigger | Pattern |
|---|---|---|
| Chief of Staff | Designed for per-tenant cron scheduling (e.g. 7 AM local time); invoked via workflow action today | Ingests unread emails + calendar events + system alerts → produces TL;DR morning briefing + critical action list |
| Assistant | Designed for a pre-meeting cadence; invoked via workflow action or on demand today | Cross-references attendees against CRM + Helpdesk → produces dossier + talking points + recent friction |
| Travel | Designed for itinerary-submitted / monitoring-cadence triggers; invoked via workflow action today | Takes an itinerary + real-time transit/weather signal → produces logistics plan + delay-risk score + suggested calendar-buffer adjustment |
4. The Concierge Audit Log
Every agent invocation creates an entry before the AI is ever called. That entry is both the deduplication key for cross-module idempotency AND the audit trail for operator review.
What's captured on every entry: - Which agent ran, and what triggered it (an order paid, a message received, a scheduled run, etc.) - The input it acted on, for full traceability - Its status: pending, completed, or failed - On success: the AI's result, an "AI Generated" flag, a human-readable impact summary, and a confidence score (0-100) - On failure: a plain-language reason — no AI provider connected, provider timed out, response failed validation, a dependent module is missing, or confidence too low - Timestamps for when it started and completed
Built-in reporting: operators can run their own reports against the audit log — grouped by agent and status over the last 24 hours, isolating an agent that's been silently failing and why, or pulling the queue of confidence-flagged items waiting for human review.
The admin UI surfaces quick-filters for AI Generated, Failed Actions, and Pending. Operators get explicit visibility into every autonomous decision the agents made — and every one they didn't make and why.